CN1213429A

CN1213429A - Method of detecting abnormality and abnormality detection system

Info

Publication number: CN1213429A
Application number: CN97192893A
Authority: CN
Inventors: 宫野尚哉
Original assignee: Sumitomo Metal Industries Ltd
Current assignee: Nippon Steel Corp
Priority date: 1996-02-16
Filing date: 1997-02-13
Publication date: 1999-04-07
Also published as: EP0881467A1; KR19990082532A; AU1671697A; EP0881467A4; TW327221B; WO1997030329A1

Abstract

A physical quantity which is emitted from a sample to be detected and which changes with the passage of time, is measured at a predetermined time interval, and abnormality of the sample is detected based on the obtained time series data. A plurality of first vectors of a suitable dimension are formed from the time series data, the first vectors are advanced in parallel for a suitable period of time to form second vectors, a deviation is calculated between the second vectors and the first vectors that are formed, the calculated deviation is compared with a preset threshold value, and whether the sample is abnormal or not is determined based upon the result of comparison.

Description

Abnormality detection method and abnormality detection system

Technical Field

The present invention relates to a method of detecting an abnormality of a manufacturing apparatus, a manufacturing plant, a manufactured product, or the like, and to a system for implementing the method.

Description of the Related Art

Sound or vibration generated by a mechanical device, concentration generated by a reaction in a chemical plant, current or voltage applied to an electronic device are measured at intervals, thereby detecting an abnormality of the mechanical device, an abnormality of reaction control, or an abnormality of product quality.

As an example of such an abnormality detection method, there is a method of detecting an abnormality by spectral analysis of a time-series signal measured at a predetermined period, which is proposed in chapters 12 to 14 of "Numerical method (Numerical receiver in C)" translated by danqing, okun, zotocten, and xiaolin (1993).

In the method, a time-series signal measured at a predetermined cycle is subjected to fast Fourier transform at appropriate frequency intervals in an appropriate frequency range to obtain a coefficient a satisfying the following formula (1)_kAnd l_kAnd find a_k ²+b_k ². At this time, at the start frequency and the end frequency of the fourier transform, in order to prevent the abrupt rise and fall of f (x) in the formula (1), an appropriate data window, such as a Hanning (Hanning) window, is set. Multiplying the Hainin window by f (x) to reduce a_k ²+b_k ²The calculation error of (2).

A to be calculated_k ²+b_k ²Plotting creates a power spectrum in the coordinate region with the horizontal axis log ω and the vertical axis log (spectral density). The periodic fluctuation component is subtracted therefrom so that an appropriate straight line is obtained by the least square method. Then, the presence or absence of an abnormality is determined based on the result of comparing the level or slope (power spectrum index) of the straight line thus obtained with a predetermined threshold value.

However, in the above-described method of detecting an abnormality by spectrum analysis, it is difficult to accurately obtain a straight line suitable for a power spectrum, and erroneous judgment occurs due to measurement noise contained in a time-series signal. In addition, it is difficult to set an appropriate data window, so the error of the power spectrum is large.

The main object of the present invention is to provide a method for detecting abnormalities and a system for implementing the same, which can accurately detect various abnormalities by using time-series data and a prediction model thereof.

Summary of the invention

The present inventors constructed a prediction model based on a plurality of past time-series signals, given the prediction model an obtained time-series signal, predicted the trend of the time-series signal, and compared the prediction result with actual time-series data to make the assumption that any abnormality can be detected.

Thus, the inventors have attempted to build many predictive models, such as: (1) linear autoregressive (e.g. "Time Series Analysis: prediction and Control" (Time Series Analysis: forms and Control) ", authors G.E.P.Box and G.M.Jenkins, publisher Holden-Day Inc., 1976), (2) neural networks (e.g." Los Alamos National laboratory report (Los Alamos National laboratory) "LA-UR 87-2662 (1987), LA-UR 88-418 (1988), authors A.S.Lapedes and R.Farber), (3) radius basis function networks (e.g." physical abstracts (physical D) "Vol.35, Vol.356 (1989), M.dagli), and (4) monomer projection methods (e.g." journal of Nature 344, Vol.734 (1990), Ma.1990), all these models are imperfect.

After trial and error, attention was paid to the time series analysis algorithm, which was proposed by Wayland et al, and published on "Physical Review Letters" volume 70, et al, 580-582 (1993), authors R.Wayland, D.Bromley, D.Pickett and A.Passeante.

As an abnormality detection method of the present invention, that is, measuring physical quantities of an object to be detected that change with time at predetermined time intervals and detecting an abnormality of the object to be detected from obtained time-series data, the method of the present invention creates a plurality of first vectors having an appropriate dimension from the time-series data, creates second vectors by shifting (transforming) the first vectors by an appropriate time interval, calculates deviations of the created first and second vectors, compares the calculated values with a predetermined threshold value, and judges whether the object to be detected is abnormal or not based on the comparison result.

In this method, the method of calculating the deviation includes the steps of: selecting an appropriate vector from the first vectors, extracting a predetermined number of vectors from the first vectors in order of distance from small to large near the selected vector, creating translation vectors by translating the extracted vectors and the selected vector by an appropriate time interval, obtaining difference vectors of differences between the created translation vectors and the selected vector and the extracted vectors, and calculating variances of the obtained difference vectors.

The abnormality detection method of the present invention is summarized below. First, a physical quantity of an object to be detected, such as a motor rotating sound or a current or voltage applied to an electronic device, is measured at an appropriate cycle, thereby obtaining time series data { x (t) }_t1 ^N(N is the number of data points). A combined embedding dimension (embedding dimension) D and a sampling time interval Δ t are set, and a vector x (t) represented by the following equation (2) is created from the time series data. It has been demonstrated that the creation of the vectors x (t) separately for the different embedding dimensions D means that the time series data are obtained separately for a plurality of physical quantities. Therefore, the abnormality can be accurately detected. The vector x (T) is translated by an appropriate time interval T Δ T to create a vector y (T) represented by the following equation (3).

X(t)={x(t),x(t-Δt),…,

x(t-(D-1)Δt)} (2)

Y(t)={x(t+TΔt),x(t+TΔt-Δt),…,

x(t+TΔt-(D-1)Δt} (3)

Selecting a suitable vector X (t (0)) from the N vectors X (t),the distances h between the selected vector X (t (0)) and all (N-1) vectors X (t) except X (t (0)) are calculated according to the following formula (4)_xAnd extracting the K nearest vectors in Euclidean (Euclidean) distance sense (K < N):

the vectors Y (t (K)) corresponding to the thus obtained (K +1) vectors X (t (K)) (K =0,1, …, K) are calculated, and then the difference vectors V (t (K)) are calculated according to equation (5). Next, when the translation error e (trans) which is the variance of the obtained (K +1) vectors V (t (K)) is obtained according to the following equation (6), the degree of clutter of the time series data is quantitatively determined as e (trans).

V(t(k))=X(t(k))-Y(t(k)) (5)

<math> <mrow> <mi>E</mi> <mrow> <mo>(</mo> <mi>trans</mi> <mo>)</mo> </mrow> <mo>=</mo> <mfrac> <mn>1</mn> <mrow> <mi>K</mi> <mo>+</mo> <mn>1</mn> </mrow> </mfrac> <munderover> <mi>Σ</mi> <mrow> <mi>k</mi> <mo>=</mo> <mn>0</mn> </mrow> <mi>K</mi> </munderover> <mfrac> <msup> <mrow> <mo>|</mo> <mi>V</mi> <mrow> <mo>(</mo> <mi>t</mi> <mrow> <mo>(</mo> <mi>k</mi> <mo>)</mo> </mrow> <mo>)</mo> </mrow> <mo>-</mo> <mo><</mo> <mi>V</mi> <mo>></mo> <mo>|</mo> </mrow> <mn>2</mn> </msup> <msup> <mrow> <mo>|</mo> <mo><</mo> <mi>V</mi> <mo>></mo> <mo>|</mo> </mrow> <mn>2</mn> </msup> </mfrac> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>6</mn> <mo>)</mo> </mrow> </mrow> </math>

Wherein,

through diligent research by the inventors, it was found that abnormalities can be easily identified. This is because the numerical amount of e (trans) increases when an abnormality occurs in which the noise ratio value increases as contained in the time system data; when an abnormality such as an increase in the power spectrum index of noise included in the time-series data occurs, the value of e (trans) becomes small, and the change in e (trans) is more significant than the power spectrum index. The present method and system have thus been invented which sets a threshold for determining the occurrence of the former anomaly and another threshold for determining the occurrence of the latter anomaly, and detects an anomaly if e (trans) exceeds the area between these two thresholds.

Brief description of the drawings

FIG. 1 is a block diagram showing the composition of an anomaly detection system according to the present invention;

FIG. 2 is a flow chart showing the detection of an abnormality by the computer of FIG. 1;

FIG. 3 is a flow chart showing the detection of an anomaly by the computer of FIG. 1;

FIG. 4 is a flow chart showing the detection of an anomaly by the computer of FIG. 1;

FIG. 5 is a flow chart showing the detection of an abnormality by the computer of FIG. 1;

FIGS. 6(A) and 6(B) are graphs showing time series data used as comparative tests;

FIG. 7 is a graph showing the result of calculating E (trans) from the two time series data of FIGS. 6(A) and 6(B) by the method of the present invention;

FIG. 8 is a graph showing the result of creating a power spectrum from the two time-series data of FIGS. 6(A) and 6(B) by a conventional method;

FIGS. 9(A) and 9(B) are graphs showing time series data used as comparative tests;

FIG. 10 is a graph showing the result of calculating E (trans) from the two time series data of FIGS. 9(A) and 9(B) by the method of the present invention;

fig. 11 is a graph showing the result of creating a power spectrum from the two time-series data of fig. 9(a) and 9(B) by a conventional method.

Description of the preferred embodiments

Embodiments of the present invention are described below with reference to the drawings.

Fig. 1 is a block diagram showing the composition of an abnormality detection system of the present invention. In fig. 1, 1 is a measuring device, such as a sonic sensor, an ultrasonic sensor, a vibration sensor, an optical sensor, a voltmeter, or an ammeter. The type of measuring device 1 is selected according to the physical quantity to be determined. The measuring apparatus 1 measures a physical quantity of an object to be measured, which is a target of abnormality detection, at a predetermined cycle. The output of the measuring device 1 is fed to an analog/digital (a/D) converter 2, the output being converted into a digital signal which is fed to a memory 32 provided to the computer 3 for storage therein as time series data.

The time-series data stored in the memory 32 is sent to the vector deviation calculating section 33 by the central processing unit 31 for a predetermined time. The vector deviation calculating unit 33 calculates e (trans) in the manner described later, and sends e (trans) to the failure diagnosing unit 34, which determines whether e (trans) is abnormal. The first and second thresholds are set at the failure condition input section 35, and the failure diagnosis section 34 compares e (trans) from the vector deviation calculation section 33 with the first and second thresholds set at the failure condition input section 35, respectively. If e (trans) is not less than the first threshold value or not more than the second threshold value, the failure diagnosing section 34 judges that e (trans) is abnormal, and causes the output section 36 to output an alarm signal to the external apparatus 4 such as a system controller or an alarm device.

Fig. 2 to 5 are flowcharts showing abnormality detection by the computer 3 in fig. 1. The vector deviation calculation unit 33 of the computer 3 sets T among the embedding dimension D, the sampling interval Δ T, and the translation time T Δ T, the vector selection number M, and the repetition numbers K and Q, and the failure condition input unit 35 sets the first and second threshold values (step S1). The memory 32 of the computer 3 stores therein the digital signal from the a/D converter 2 as time-series data (step S2), and the time-series data stored in the memory 32 is sent to the vector deviation calculating section 33 via the central processing unit 31 for a predetermined time.

According to known time series data { X (t) }_t-1 ^N(N is the number of data points), the vector deviation calculation unit 33 creates a vector represented by the following expression (2) (step S3).

X(t)={x(t),x(t-Δt),…,

x(t-(D-1)Δt} (2)

Further, the vector deviation calculation section 33 translates the vector by an appropriate time interval T Δ T and creates a vector y (T) represented by the following expression (3) (step S4).

Y(t)={x(t+TΔt),x(t+TΔt-Δt),…,

x(t+TΔt-(D-1)Δt)} (3)

The vector deviation calculation section 33 generates an arbitrary random number (step S5), and selects M (M < N) vectors X (t (0)) from the N vectors X (t) created at S3 by using the generated random number (step S6), and then appropriately specifies one vector X (t (0)) (step S7). The vector deviation calculation unit 33 calculates a predetermined vector X (t (0)) and a division vector according to the following expression (4)Respective distances h between the remaining (N-1) vectors X (t) other than X (t (0)_x(step S8), and rearranges these calculated distances in ascending order, and extracts K (K < M) nearest vectors from the head, thereby obtaining a vector X (t (K)) (K =1, …, K) (step S9).

By shifting the vector X (T (K)) (K =0,1, …, K) obtained at S7 and S9 by the time interval T Δ T, the vector deviation calculation section 33 creates the vector Y (T (K)) thus obtained (step S10).

The vector deviation calculation section 33 calculates (k +1) difference vectors V (t (k)) according to the following expression (5) (step S11), and then obtains an original (primary) translational error e (trans)) from each difference vector V (t (k)) and the following expression (7) (step S12).

V(t(k))=X(t(k))-Y(t(k)) (5)

<math> <mrow> <mi>E</mi> <msub> <mrow> <mo>(</mo> <mi>trans</mi> <mo>)</mo> </mrow> <mn>1</mn> </msub> <mo>=</mo> <mfrac> <mn>1</mn> <mrow> <mi>K</mi> <mo>+</mo> <mn>1</mn> </mrow> </mfrac> <munderover> <mi>Σ</mi> <mrow> <mi>k</mi> <mo>-</mo> <mn>0</mn> </mrow> <mi>K</mi> </munderover> <mfrac> <msup> <mrow> <mo>|</mo> <mi>V</mi> <mrow> <mo>(</mo> <mi>t</mi> <mrow> <mo>(</mo> <mi>k</mi> <mo>)</mo> </mrow> <mo>)</mo> </mrow> <mo>-</mo> <mo><</mo> <mi>V</mi> <mo>></mo> <mo>|</mo> </mrow> <mn>2</mn> </msup> <msup> <mrow> <mo>|</mo> <mo><</mo> <mi>V</mi> <mo>></mo> <mo>|</mo> </mrow> <mn>2</mn> </msup> </mfrac> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>6</mn> <mo>)</mo> </mrow> </mrow> </math>

Wherein,

the vector deviation calculation section 33 repeats steps S7 to S12 until it judges at step S13 that M original translational errors E (trans) have been obtained₁And is the original translation error E (trans) for the M vectors X (t (0)) selected at step S6₁. When the vector deviation calculating part 33 judges that M original translation errors E (trans) are obtained₁It calculates its median value to obtain the original mean translational error E (trans)₁₁(step S14).

Next, the vector deviation calculation section 33 repeats steps S5 to S14 until it judges at step S15 that Q original average translational errors E (trans) are obtained₁₁Then, when it is judged that Q original average translation errors E (trans) have been obtained₁₁It calculates its average value to obtain e (trans) (step 16). As a result, e (trans) of the time series data can be accurately obtained.

When e (trans) is obtained, the vector deviation calculation section 33 sends it to the failure diagnosis section 34. The failure diagnosis unit 34 compares e (trans) from the vector deviation calculation unit 33 with the first threshold value and/or the second threshold value set in the failure condition input unit 35 (step 17). When e (trans) is not less than the first threshold value or not more than the second threshold value, the failure diagnosing section 34 judges e (trans) as abnormal (step 18), and outputs an alarm signal from the output section 36 to the external apparatus 4 (step 4).

Next, the results of comparison for detecting abnormality by two methods, one using the method of the present invention and the other using a conventional method, are described below.

Fig. 6 and 9 are graphs showing time series data used for comparative tests. The vertical axis represents signal strength and the horizontal axis represents time from the start of measurement. Fig. 6 is time-series data representing vibration of the motor rotating portion of the simulated fan: FIG. 6(A) shows time-series data when no abnormality occurs; fig. 6(B) shows time-series data at the time of occurrence of an abnormality. The signal intensity represented by the vertical axis is obtained by converting the vibration into an electric signal with a vibration sensor, but a reflected wave of the vibrating portion may also be detected with an optical sensor. Further, the vibration component can also be obtained by a voltage value obtained by measuring a voltage characteristic of the motor with a voltmeter or a current value obtained by measuring a current characteristic of the motor with a ammeter.

Fig. 9 is a graph showing time-series data of an analog audio signal: FIG. 9(A) shows time-series data when no abnormality occurs; fig. 9(B) shows time-series data when an abnormality occurs. The signal intensity represented by the vertical axis is obtained by converting sound into an electric signal with an acoustic wave sensor.

FIG. 7 is a graph showing the result of calculating E (trans) from the two time-series data of FIGS. 6(A) and 6(B) by the method of the present invention. ● denotes non-exception. And o indicates an abnormality. Here, the sampling interval Δ T is set to 1, T of the translation time T Δ T is set to 5, the vector selection number M is set to 300, the vector selection number K is set to 4, the repetition number Q is selected to 20, and the embedding dimension D is set to 5, 6, 7, 8, 9, and 10. Further, fig. 8 is a graph showing the result of creating a power spectrum from the two kinds of time-series data shown in fig. 6(a) and 6(B) by a conventional method. The solid line represents non-anomalies and the dashed line represents anomalies.

As is clear from fig. 7, in the method of the present invention, e (trans) in the abnormal case is clearly separated from e (trans) in the non-abnormal case, and by setting 0.05 as a threshold, it is possible to accurately determine whether an abnormality or non-abnormality is present. Meanwhile, as is clear from fig. 8, in the conventional method, many portions of the power spectrum in the abnormal case overlap with the power spectrum in the non-abnormal case, and thus a judgment error occurs.

Fig. 10 is a graph showing the result of calculating e (trans) from the two types of time-series data of fig. 9(a) and 9(B) by the present invention. ● indicates no anomaly and O indicates an anomaly. Here, the sampling interval Δ T is set to 1, T of the translation time T Δ T is set to 5, the vector selection number M is set to 300, the vector selection number K is set to 4, the repetition number Q is set to 20, and the embedding dimension D is set to 11, 12, 13, 14, and 15. Further, fig. 11 is a graph showing the result of creating a power spectrum from the two kinds of time-series data shown in fig. 9(a) and 9(B) by a conventional method. The solid line represents non-anomalies and the dashed line represents anomalies.

As is clear from fig. 10, in the method of the present invention, e (trans) in the abnormal case is clearly separated from e (trans) in the non-abnormal case, and by setting 0.90 as a threshold, it is possible to accurately determine whether an abnormality or non-abnormality is present. Meanwhile, as is clear from fig. 11, in the conventional method, many portions of the power spectrum in the abnormal case overlap with the power spectrum in the non-abnormal case, and thus a judgment error occurs. Industrial applicability

As described in detail above, in the present invention, even if there is a change in the physical quantity to be measured due to the nonlinear dynamics, various abnormalities can be accurately detected. Therefore, an excellent effect is produced such that the operation of the manufacturing plant is accurately managed, the parts of the manufacturing apparatus are accurately replaced, and the quality of the product is accurately managed.

Claims

1. An abnormality detection method that measures a physical quantity of an object to be detected that changes with time at predetermined time intervals and detects an abnormality of the object to be detected based on measured time-series data, the method comprising the steps of:

creating a plurality of first vectors of appropriate dimensions from the time series data;

creating second vectors by translating said first vectors by an appropriate time interval;

calculating a deviation of said created first and second vectors;

comparing said calculated values to a predetermined threshold; and

and judging whether the object to be detected is abnormal or not according to the comparison result.

2. The anomaly detection method in accordance with claim 1, wherein said step of calculating said deviations comprises:

a first step of selecting an appropriate vector from said first vectors;

a second step of extracting a predetermined number of vectors from the first vectors in order of a distance from small to large approaching the selected vector;

a third step of creating translation vectors by translating said extracted vectors and said selected vector by an appropriate time interval;

a fourth step of obtaining difference vectors between said created translation vectors and said selected vector and said extracted vectors; and

and a fifth step of calculating the variance of the obtained difference vectors.

3. The abnormality detection method according to claim 2, further comprising the steps of:

repeating said first to fifth steps while changing said selected vector, obtaining a plurality of intermediate values of the means by repeatedly calculating a plurality of variances; and

obtaining a median or average of the calculated plurality of variances, and setting the average of the obtained plurality of median or average as a calculated value to be compared with the threshold.

4. The abnormality detection method according to claim 2, further comprising the steps of:

calculating a plurality of variances by repeating the first to fifth steps while permuting the selected vector;

obtaining a median or mean of the calculated plurality of variances; and

the obtained median or average value is set as a calculated value to be compared with the threshold value.

5. An abnormality detection system that measures a physical quantity of an object to be detected that changes with time at predetermined time intervals and detects an abnormality of the object to be detected based on measured time-series data, the system comprising:

means for creating a plurality of first vectors of appropriate dimensions from the time series data;

means for creating second vectors by translating said first vector by an appropriate time interval;

calculating means for calculating a deviation between the first and second vectors that have been created;

means for comparing said calculated values with predetermined thresholds; and

and judging whether the object to be detected is abnormal according to the comparison result.

6. The anomaly detection system according to claim 5, wherein said calculation means comprises:

means for selecting an appropriate vector from said first vectors;

means for extracting a predetermined number of vectors from said first vectors in order of distance from small to large near the selected vector;

means for creating translation vectors by translating said extracted vectors and said selected vector for an appropriate time interval;

means for obtaining difference vectors between said created translation vectors and said selected vector and said extracted vectors; and

means for calculating the variance of said obtained difference vectors.

7. An anomaly detection system for detecting anomalies in an object to be detected, the system comprising:

a measuring device that measures a physical quantity of an object to be detected that changes with time at predetermined time intervals;

means for storing time series data of measurement values obtained by the measuring means;

means for creating a plurality of first vectors of appropriate dimensions from said stored time series data;

means for creating second vectors by translating said first vectors by an appropriate time interval;

means for calculating a deviation between said created first and second vectors;

means for comparing said calculated value with a predetermined threshold value; and

and detecting the object to be detected as abnormal according to the comparison result.

8. The anomaly detection system according to claim 7, wherein said measuring means is a device selected from the group consisting of: acoustic wave sensors, ultrasonic sensors, vibration sensors, optical sensors, voltmeters, and ammeters.